[Feature]: Support FCN,DeeepLabV3, DeepLabV3Plus in mmseg with ONNXRuntime and TensorRT (#31)

* fix mask empty result

* support fcn exporting to ONNX for ort and trt in whole mode

* resolve comments

* remove unnecessary code

* update prepare_input

* rewrite psp_head & aspp_head

* test fcn deeplabv3 deeplabv3plus with trt

.isort.cfg

@@ -1,2 +1,2 @@
 [settings]
-known_third_party = mmcls,mmcv,mmdet,numpy,onnx,packaging,pyppl,pytest,setuptools,tensorrt,torch
+known_third_party = mmcls,mmcv,mmdet,mmseg,numpy,onnx,packaging,pyppl,pytest,setuptools,tensorrt,torch

configs/mmseg/base.py

@@ -0,0 +1,18 @@
+_base_ = ['../_base_/torch2onnx.py']
+codebase = 'mmseg'
+pytorch2onnx = dict(
+    input_names=['input'],
+    output_names=['output'],
+    dynamic_axes={
+        'input': {
+            0: 'batch',
+            2: 'height',
+            3: 'width'
+        },
+        'output': {
+            0: 'batch',
+            2: 'height',
+            3: 'width'
+        },
+    },
+)

configs/mmseg/onnxruntime.py

@@ -0,0 +1 @@
+_base_ = ['./base.py', '../_base_/backends/onnxruntime.py']

configs/mmseg/tensorrt.py

@@ -0,0 +1,7 @@
+_base_ = ['./base.py', '../_base_/backends/tensorrt.py']
+tensorrt_params = dict(model_params=[
+    dict(
+        opt_shape_dict=dict(
+            input=[[1, 3, 512, 512], [1, 3, 1024, 2048], [1, 3, 2048, 2048]]),
+        max_workspace_size=1 << 30)
+])

mmdeploy/apis/test.py

@@ -10,13 +10,13 @@ from mmdeploy.apis.utils import assert_module_exist
 
 
 def prepare_data_loader(codebase: str, model_cfg: Union[str, mmcv.Config]):
+    assert_module_exist(codebase)
     # load model_cfg if necessary
     if isinstance(model_cfg, str):
         model_cfg = mmcv.Config.fromfile(model_cfg)
 
     if codebase == 'mmcls':
         from mmcls.datasets import (build_dataloader, build_dataset)
-        assert_module_exist(codebase)
         # build dataset and dataloader
         dataset = build_dataset(model_cfg.data.test)
         data_loader = build_dataloader(
@@ -27,7 +27,6 @@ def prepare_data_loader(codebase: str, model_cfg: Union[str, mmcv.Config]):
             round_up=False)
 
     elif codebase == 'mmdet':
-        assert_module_exist(codebase)
         from mmdet.datasets import (build_dataloader, build_dataset,
                                     replace_ImageToTensor)
         # in case the test dataset is concatenated
@@ -58,7 +57,17 @@ def prepare_data_loader(codebase: str, model_cfg: Union[str, mmcv.Config]):
             workers_per_gpu=model_cfg.data.workers_per_gpu,
             dist=False,
             shuffle=False)
-
+    elif codebase == 'mmseg':
+        from mmseg.datasets import build_dataset, build_dataloader
+        model_cfg.data.test.test_mode = True
+        dataset = build_dataset(model_cfg.data.test)
+        samples_per_gpu = 1
+        data_loader = build_dataloader(
+            dataset,
+            samples_per_gpu=samples_per_gpu,
+            workers_per_gpu=model_cfg.data.workers_per_gpu,
+            dist=False,
+            shuffle=False)
     else:
         raise NotImplementedError(f'Unknown codebase type: {codebase}')
 
@@ -71,16 +80,18 @@ def single_gpu_test(codebase: str,
                     show: bool = False,
                     out_dir: Any = None,
                     show_score_thr: float = 0.3):
+    assert_module_exist(codebase)
+
     if codebase == 'mmcls':
-        assert_module_exist(codebase)
         from mmcls.apis import single_gpu_test
         outputs = single_gpu_test(model, data_loader, show, out_dir)
     elif codebase == 'mmdet':
-        assert_module_exist(codebase)
         from mmdet.apis import single_gpu_test
         outputs = single_gpu_test(model, data_loader, show, out_dir,
                                   show_score_thr)
-
+    elif codebase == 'mmseg':
+        from mmseg.apis import single_gpu_test
+        outputs = single_gpu_test(model, data_loader, show, out_dir)
     else:
         raise NotImplementedError(f'Unknown codebase type: {codebase}')
     return outputs
@@ -138,5 +149,15 @@ def post_process_outputs(outputs,
             eval_kwargs.update(dict(metric=metrics, **kwargs))
             print(dataset.evaluate(outputs, **eval_kwargs))
 
+    elif codebase == 'mmseg':
+        if out:
+            print(f'\nwriting results to {out}')
+            mmcv.dump(outputs, out)
+        kwargs = {} if metric_options is None else metric_options
+        if format_only:
+            dataset.format_results(outputs, **kwargs)
+        if metrics:
+            dataset.evaluate(outputs, metrics, **kwargs)
+
     else:
         raise NotImplementedError(f'Unknown codebase type: {codebase}')

mmdeploy/apis/utils.py

@@ -31,49 +31,50 @@ def init_model(codebase: str,
                model_checkpoint: Optional[str] = None,
                device: str = 'cuda:0',
                cfg_options: Optional[Dict] = None):
-    # mmcls
+    assert_module_exist(codebase)
     if codebase == 'mmcls':
-        assert_module_exist(codebase)
         from mmcls.apis import init_model
         model = init_model(model_cfg, model_checkpoint, device, cfg_options)
 
     elif codebase == 'mmdet':
-        assert_module_exist(codebase)
         from mmdet.apis import init_detector
         model = init_detector(model_cfg, model_checkpoint, device, cfg_options)
 
     elif codebase == 'mmseg':
-        assert_module_exist(codebase)
         from mmseg.apis import init_segmentor
+        from mmdeploy.mmseg.export import convert_syncbatchnorm
         model = init_segmentor(model_cfg, model_checkpoint, device)
+        model = convert_syncbatchnorm(model)
 
     else:
         raise NotImplementedError(f'Unknown codebase type: {codebase}')
 
-    return model
+    return model.eval()
 
 
 def create_input(codebase: str,
                  model_cfg: Union[str, mmcv.Config],
                  imgs: Any,
                  device: str = 'cuda:0'):
+    assert_module_exist(codebase)
     if isinstance(model_cfg, str):
         model_cfg = mmcv.Config.fromfile(model_cfg)
     elif not isinstance(model_cfg, (mmcv.Config, mmcv.ConfigDict)):
         raise TypeError('config must be a filename or Config object, '
                         f'but got {type(model_cfg)}')
-
     cfg = model_cfg.copy()
     if codebase == 'mmcls':
-        assert_module_exist(codebase)
         from mmdeploy.mmcls.export import create_input
         return create_input(cfg, imgs, device)
 
     elif codebase == 'mmdet':
-        assert_module_exist(codebase)
         from mmdeploy.mmdet.export import create_input
         return create_input(cfg, imgs, device)
 
+    elif codebase == 'mmseg':
+        from mmdeploy.mmseg.export import create_input
+        return create_input(cfg, imgs, device)
+
     else:
         raise NotImplementedError(f'Unknown codebase type: {codebase}')
 
@@ -94,8 +95,8 @@ def init_backend_model(model_files: Sequence[str],
                        backend: str,
                        class_names: Sequence[str],
                        device_id: int = 0):
+    assert_module_exist(codebase)
     if codebase == 'mmcls':
-        assert_module_exist(codebase)
         if backend == 'onnxruntime':
             from mmdeploy.mmcls.export import ONNXRuntimeClassifier
             backend_model = ONNXRuntimeClassifier(
@@ -120,7 +121,6 @@ def init_backend_model(model_files: Sequence[str],
         return backend_model
 
     elif codebase == 'mmdet':
-        assert_module_exist(codebase)
         if backend == 'onnxruntime':
             from mmdeploy.mmdet.export import ONNXRuntimeDetector
             backend_model = ONNXRuntimeDetector(
@@ -137,61 +137,55 @@ def init_backend_model(model_files: Sequence[str],
             raise NotImplementedError(f'Unsupported backend type: {backend}')
         return backend_model
 
+    elif codebase == 'mmseg':
+        if backend == 'onnxruntime':
+            from mmdeploy.mmseg.export import ONNXRuntimeSegmentor
+            backend_model = ONNXRuntimeSegmentor(
+                model_files[0], class_names=class_names, device_id=device_id)
+        elif backend == 'tensorrt':
+            from mmdeploy.mmseg.export import TensorRTSegmentor
+            backend_model = TensorRTSegmentor(
+                model_files[0], class_names=class_names, device_id=device_id)
+        else:
+            raise NotImplementedError(f'Unsupported backend type: {backend}')
+        return backend_model
     else:
         raise NotImplementedError(f'Unknown codebase type: {codebase}')
 
 
 def get_classes_from_config(codebase: str, model_cfg: Union[str, mmcv.Config]):
+    assert_module_exist(codebase)
+
     model_cfg_str = model_cfg
+    if isinstance(model_cfg, str):
+        model_cfg = mmcv.Config.fromfile(model_cfg)
+    elif not isinstance(model_cfg, (mmcv.Config, mmcv.ConfigDict)):
+        raise TypeError('config must be a filename or Config object, '
+                        f'but got {type(model_cfg)}')
+
     if codebase == 'mmcls':
-        assert_module_exist(codebase)
-        if isinstance(model_cfg, str):
-            model_cfg = mmcv.Config.fromfile(model_cfg)
-        elif not isinstance(model_cfg, (mmcv.Config, mmcv.ConfigDict)):
-            raise TypeError('config must be a filename or Config object, '
-                            f'but got {type(model_cfg)}')
-
         from mmcls.datasets import DATASETS
-        module_dict = DATASETS.module_dict
-        data_cfg = model_cfg.data
-
-        if 'train' in data_cfg:
-            module = module_dict[data_cfg.train.type]
-        elif 'val' in data_cfg:
-            module = module_dict[data_cfg.val.type]
-        elif 'test' in data_cfg:
-            module = module_dict[data_cfg.test.type]
-        else:
-            raise RuntimeError(f'No dataset config found in: {model_cfg_str}')
-
-        return module.CLASSES
-
-    if codebase == 'mmdet':
-        assert_module_exist(codebase)
-        if isinstance(model_cfg, str):
-            model_cfg = mmcv.Config.fromfile(model_cfg)
-        elif not isinstance(model_cfg, (mmcv.Config, mmcv.ConfigDict)):
-            raise TypeError('config must be a filename or Config object, '
-                            f'but got {type(model_cfg)}')
-
+    elif codebase == 'mmdet':
         from mmdet.datasets import DATASETS
-        module_dict = DATASETS.module_dict
-        data_cfg = model_cfg.data
-
-        if 'train' in data_cfg:
-            module = module_dict[data_cfg.train.type]
-        elif 'val' in data_cfg:
-            module = module_dict[data_cfg.val.type]
-        elif 'test' in data_cfg:
-            module = module_dict[data_cfg.test.type]
-        else:
-            raise RuntimeError(f'No dataset config found in: {model_cfg_str}')
-
-        return module.CLASSES
-
+    elif codebase == 'mmseg':
+        from mmseg.datasets import DATASETS
     else:
         raise NotImplementedError(f'Unknown codebase type: {codebase}')
 
+    module_dict = DATASETS.module_dict
+    data_cfg = model_cfg.data
+
+    if 'train' in data_cfg:
+        module = module_dict[data_cfg.train.type]
+    elif 'val' in data_cfg:
+        module = module_dict[data_cfg.val.type]
+    elif 'test' in data_cfg:
+        module = module_dict[data_cfg.test.type]
+    else:
+        raise RuntimeError(f'No dataset config found in: {model_cfg_str}')
+
+    return module.CLASSES
+
 
 def check_model_outputs(codebase: str,
                         image: Union[str, np.ndarray],
@@ -199,10 +193,12 @@ def check_model_outputs(codebase: str,
                         model,
                         output_file: str,
                         backend: str,
+                        dataset: str = None,
                         show_result=False):
+    assert_module_exist(codebase)
     show_img = mmcv.imread(image) if isinstance(image, str) else image
+
     if codebase == 'mmcls':
-        assert_module_exist(codebase)
         output_file = None if show_result else output_file
         with torch.no_grad():
             scores = model(**model_inputs, return_loss=False)[0]
@@ -221,7 +217,6 @@ def check_model_outputs(codebase: str,
                 out_file=output_file)
 
     elif codebase == 'mmdet':
-        assert_module_exist(codebase)
         output_file = None if show_result else output_file
         score_thr = 0.3
         with torch.no_grad():
@@ -234,5 +229,20 @@ def check_model_outputs(codebase: str,
                 win_name=backend,
                 out_file=output_file)
 
+    elif codebase == 'mmseg':
+        output_file = None if show_result else output_file
+        from mmseg.core.evaluation import get_palette
+        dataset = 'cityscapes' if dataset is None else dataset
+        palette = get_palette(dataset)
+        with torch.no_grad():
+            results = model(**model_inputs, return_loss=False, rescale=True)
+            model.show_result(
+                show_img,
+                results,
+                palette=palette,
+                show=True,
+                win_name=backend,
+                out_file=output_file,
+                opacity=0.5)
     else:
         raise NotImplementedError(f'Unknown codebase type: {codebase}')

mmdeploy/mmdet/export/model_wrappers.py

@@ -74,7 +74,8 @@ class DeployBaseDetector(BaseDetector):
                 img_h, img_w = img_metas[i]['img_shape'][:2]
                 ori_h, ori_w = img_metas[i]['ori_shape'][:2]
                 masks = masks[:, :img_h, :img_w]
-                if rescale and batch_masks.shape[1] > 0:
+                # avoid to resize masks with zero dim
+                if rescale and masks.shape[0] != 0:
                     masks = masks.astype(np.float32)
                     masks = torch.from_numpy(masks)
                     masks = torch.nn.functional.interpolate(

mmdeploy/mmseg/__init__.py

@@ -0,0 +1,2 @@
+from .export import *  # noqa: F401,F403
+from .models import *  # noqa: F401,F403

mmdeploy/mmseg/export/__init__.py

@@ -0,0 +1,8 @@
+from .model_wrappers import ONNXRuntimeSegmentor, TensorRTSegmentor
+from .onnx_helper import convert_syncbatchnorm
+from .prepare_input import create_input
+
+__all__ = [
+    'create_input', 'ONNXRuntimeSegmentor', 'TensorRTSegmentor',
+    'convert_syncbatchnorm'
+]

mmdeploy/mmseg/export/model_wrappers.py

@@ -0,0 +1,117 @@
+import os.path as osp
+import warnings
+from typing import Sequence
+
+import numpy as np
+import torch
+from mmseg.models.segmentors.base import BaseSegmentor
+from mmseg.ops import resize
+
+
+class DeployBaseSegmentor(BaseSegmentor):
+
+    def __init__(self, class_names: Sequence[str], device_id: int):
+        super(DeployBaseSegmentor, self).__init__(init_cfg=None)
+        self.CLASSES = class_names
+        self.device_id = device_id
+        self.PALETTE = None
+
+    def extract_feat(self, imgs):
+        raise NotImplementedError('This method is not implemented.')
+
+    def encode_decode(self, img, img_metas):
+        raise NotImplementedError('This method is not implemented.')
+
+    def forward_train(self, imgs, img_metas, **kwargs):
+        raise NotImplementedError('This method is not implemented.')
+
+    def simple_test(self, img, img_meta, **kwargs):
+        raise NotImplementedError('This method is not implemented.')
+
+    def aug_test(self, imgs, img_metas, **kwargs):
+        raise NotImplementedError('This method is not implemented.')
+
+    def forward(self, img, img_metas, **kwargs):
+        seg_pred = self.forward_test(img, img_metas, **kwargs)
+        # whole mode supports dynamic shape
+        ori_shape = img_metas[0][0]['ori_shape']
+        if not (ori_shape[0] == seg_pred.shape[-2]
+                and ori_shape[1] == seg_pred.shape[-1]):
+            seg_pred = torch.from_numpy(seg_pred).float()
+            seg_pred = resize(
+                seg_pred, size=tuple(ori_shape[:2]), mode='nearest')
+            seg_pred = seg_pred.long().detach().cpu().numpy()
+        # remove unnecessary dim
+        seg_pred = seg_pred.squeeze(1)
+        seg_pred = list(seg_pred)
+        return seg_pred
+
+
+class ONNXRuntimeSegmentor(DeployBaseSegmentor):
+
+    def __init__(self, onnx_file: str, class_names: Sequence[str],
+                 device_id: int):
+        super(ONNXRuntimeSegmentor, self).__init__(class_names, device_id)
+
+        import onnxruntime as ort
+        from mmdeploy.apis.onnxruntime import get_ops_path
+
+        # get the custom op path
+        ort_custom_op_path = get_ops_path()
+        session_options = ort.SessionOptions()
+        # register custom op for onnxruntime
+        if osp.exists(ort_custom_op_path):
+            session_options.register_custom_ops_library(ort_custom_op_path)
+        sess = ort.InferenceSession(onnx_file, session_options)
+        providers = ['CPUExecutionProvider']
+        options = [{}]
+        is_cuda_available = ort.get_device() == 'GPU'
+        if is_cuda_available:
+            providers.insert(0, 'CUDAExecutionProvider')
+            options.insert(0, {'device_id': device_id})
+
+        sess.set_providers(providers, options)
+
+        self.sess = sess
+        self.io_binding = sess.io_binding()
+        self.output_names = [_.name for _ in sess.get_outputs()]
+        for name in self.output_names:
+            self.io_binding.bind_output(name)
+
+    def forward_test(self, imgs, img_metas, **kwargs):
+        input_data = imgs[0]
+        device_type = input_data.device.type
+        self.io_binding.bind_input(
+            name='input',
+            device_type=device_type,
+            device_id=self.device_id,
+            element_type=np.float32,
+            shape=input_data.shape,
+            buffer_ptr=input_data.data_ptr())
+        self.sess.run_with_iobinding(self.io_binding)
+        seg_pred = self.io_binding.copy_outputs_to_cpu()[0]
+        return seg_pred
+
+
+class TensorRTSegmentor(DeployBaseSegmentor):
+
+    def __init__(self, trt_file: str, class_names: Sequence[str],
+                 device_id: int):
+        super(TensorRTSegmentor, self).__init__(class_names, device_id)
+
+        from mmdeploy.apis.tensorrt import TRTWrapper, load_tensorrt_plugin
+        try:
+            load_tensorrt_plugin()
+        except (ImportError, ModuleNotFoundError):
+            warnings.warn('If input model has custom plugins, \
+                you may have to build backend ops with TensorRT')
+        model = TRTWrapper(trt_file)
+        self.model = model
+        self.output_name = self.model.output_names[0]
+
+    def forward_test(self, imgs, img_metas, **kwargs):
+        input_data = imgs[0].contiguous()
+        with torch.cuda.device(self.device_id), torch.no_grad():
+            seg_pred = self.model({'input': input_data})[self.output_name]
+        seg_pred = seg_pred.detach().cpu().numpy()
+        return seg_pred

mmdeploy/mmseg/export/onnx_helper.py

@@ -0,0 +1,22 @@
+import torch
+
+
+def convert_syncbatchnorm(module):
+    module_output = module
+    if isinstance(module, torch.nn.SyncBatchNorm):
+        module_output = torch.nn.BatchNorm2d(module.num_features, module.eps,
+                                             module.momentum, module.affine,
+                                             module.track_running_stats)
+        if module.affine:
+            module_output.weight.data = module.weight.data.clone().detach()
+            module_output.bias.data = module.bias.data.clone().detach()
+            # keep requires_grad unchanged
+            module_output.weight.requires_grad = module.weight.requires_grad
+            module_output.bias.requires_grad = module.bias.requires_grad
+        module_output.running_mean = module.running_mean
+        module_output.running_var = module.running_var
+        module_output.num_batches_tracked = module.num_batches_tracked
+    for name, child in module.named_children():
+        module_output.add_module(name, convert_syncbatchnorm(child))
+    del module
+    return module_output

mmdeploy/mmseg/export/prepare_input.py

@@ -0,0 +1,47 @@
+from typing import Any, Union
+
+import mmcv
+import numpy as np
+from mmcv.parallel import collate, scatter
+from mmseg.apis.inference import LoadImage
+from mmseg.datasets.pipelines import Compose
+
+
+def create_input(model_cfg: Union[str, mmcv.Config],
+                 imgs: Any,
+                 device: str = 'cuda:0'):
+    if isinstance(model_cfg, str):
+        model_cfg = mmcv.Config.fromfile(model_cfg)
+    elif not isinstance(model_cfg, (mmcv.Config, mmcv.ConfigDict)):
+        raise TypeError('config must be a filename or Config object, '
+                        f'but got {type(model_cfg)}')
+    cfg = model_cfg.copy()
+
+    if not isinstance(imgs, (list, tuple)):
+        imgs = [imgs]
+
+    if isinstance(imgs[0], np.ndarray):
+        cfg = cfg.copy()
+        # set loading pipeline type
+        cfg.data.test.pipeline[0].type = 'LoadImageFromWebcam'
+
+    cfg.data.test.pipeline[1]['transforms'][0]['keep_ratio'] = False
+    cfg.data.test.pipeline = [LoadImage()] + cfg.data.test.pipeline[1:]
+
+    test_pipeline = Compose(cfg.data.test.pipeline)
+    datas = []
+    for img in imgs:
+        # prepare data
+        data = dict(img=img)
+        # build the data pipeline
+        data = test_pipeline(data)
+        datas.append(data)
+
+    data = collate(datas, samples_per_gpu=len(imgs))
+
+    data['img_metas'] = [img_metas.data[0] for img_metas in data['img_metas']]
+    data['img'] = [img.data[0][None, :] for img in data['img']]
+    if device != 'cpu':
+        data = scatter(data, [device])[0]
+
+    return data, data['img']

mmdeploy/mmseg/models/__init__.py

@@ -0,0 +1,2 @@
+from .decode_heads import *  # noqa: F401,F403
+from .segmentors import *  # noqa: F401,F403

mmdeploy/mmseg/models/decode_heads/__init__.py

@@ -0,0 +1,4 @@
+from .aspp_head import forward_of_aspp_head
+from .psp_head import forward_of_ppm
+
+__all__ = ['forward_of_aspp_head', 'forward_of_ppm']

mmdeploy/mmseg/models/decode_heads/aspp_head.py

@@ -0,0 +1,30 @@
+import torch
+from mmseg.ops import resize
+
+from mmdeploy.core import FUNCTION_REWRITER
+from mmdeploy.utils import is_dynamic_shape
+
+
+@FUNCTION_REWRITER.register_rewriter(
+    func_name='mmseg.models.decode_heads.ASPPHead.forward')
+def forward_of_aspp_head(ctx, self, inputs):
+    x = self._transform_inputs(inputs)
+    deploy_cfg = ctx.cfg
+    is_dynamic_flag = is_dynamic_shape(deploy_cfg)
+    # get origin input shape as tensor to support onnx dynamic shape
+    size = x.shape[2:]
+    if not is_dynamic_flag:
+        size = [int(val) for val in size]
+
+    aspp_outs = [
+        resize(
+            self.image_pool(x),
+            size=size,
+            mode='bilinear',
+            align_corners=self.align_corners)
+    ]
+    aspp_outs.extend(self.aspp_modules(x))
+    aspp_outs = torch.cat(aspp_outs, dim=1)
+    output = self.bottleneck(aspp_outs)
+    output = self.cls_seg(output)
+    return output

mmdeploy/mmseg/models/decode_heads/psp_head.py

@@ -0,0 +1,26 @@
+from mmseg.ops import resize
+
+from mmdeploy.core import FUNCTION_REWRITER
+from mmdeploy.utils import is_dynamic_shape
+
+
+@FUNCTION_REWRITER.register_rewriter(
+    func_name='mmseg.models.decode_heads.psp_head.PPM.forward')
+def forward_of_ppm(ctx, self, x):
+    deploy_cfg = ctx.cfg
+    is_dynamic_flag = is_dynamic_shape(deploy_cfg)
+    # get origin input shape as tensor to support onnx dynamic shape
+    size = x.shape[2:]
+    if not is_dynamic_flag:
+        size = [int(val) for val in size]
+
+    ppm_outs = []
+    for ppm in self:
+        ppm_out = ppm(x)
+        upsampled_ppm_out = resize(
+            ppm_out,
+            size=size,
+            mode='bilinear',
+            align_corners=self.align_corners)
+        ppm_outs.append(upsampled_ppm_out)
+    return ppm_outs

mmdeploy/mmseg/models/segmentors/__init__.py

@@ -0,0 +1,4 @@
+from .base import forward_of_base_segmentor
+from .encoder_decoder import simple_test_of_encoder_decoder
+
+__all__ = ['forward_of_base_segmentor', 'simple_test_of_encoder_decoder']

mmdeploy/mmseg/models/segmentors/base.py

@@ -0,0 +1,22 @@
+import torch
+
+from mmdeploy.core import FUNCTION_REWRITER
+from mmdeploy.utils import is_dynamic_shape
+
+
+@FUNCTION_REWRITER.register_rewriter(
+    func_name='mmseg.models.segmentors.BaseSegmentor.forward')
+def forward_of_base_segmentor(ctx, self, img, img_metas=None, **kwargs):
+    if img_metas is None:
+        img_metas = {}
+    assert isinstance(img_metas, dict)
+    assert isinstance(img, torch.Tensor)
+
+    deploy_cfg = ctx.cfg
+    is_dynamic_flag = is_dynamic_shape(deploy_cfg)
+    # get origin input shape as tensor to support onnx dynamic shape
+    img_shape = img.shape[2:]
+    if not is_dynamic_flag:
+        img_shape = [int(val) for val in img_shape]
+    img_metas['img_shape'] = img_shape
+    return self.simple_test(img, img_metas, **kwargs)

mmdeploy/mmseg/models/segmentors/encoder_decoder.py

@@ -0,0 +1,21 @@
+import torch.nn.functional as F
+from mmseg.ops import resize
+
+from mmdeploy.core import FUNCTION_REWRITER
+
+
+@FUNCTION_REWRITER.register_rewriter(
+    func_name='mmseg.models.segmentors.EncoderDecoder.simple_test')
+def simple_test_of_encoder_decoder(ctx, self, img, img_meta, **kwargs):
+    x = self.extract_feat(img)
+    seg_logit = self._decode_head_forward_test(x, img_meta)
+    seg_logit = resize(
+        input=seg_logit,
+        size=img_meta['img_shape'],
+        mode='bilinear',
+        align_corners=self.align_corners)
+    seg_logit = F.softmax(seg_logit, dim=1)
+    seg_pred = seg_logit.argmax(dim=1)
+    # our inference backend only support 4D output
+    seg_pred = seg_pred.unsqueeze(1)
+    return seg_pred